← All stories
● Covered by 1 source · 1 reportMedium impact1 neutral

Tufa Labs' tufazip Compresses 1GB Wikipedia Data to 105MB Losslessly with Self-Learning AI

🔄 Updated 8h ago
New to BrevFeed? We gather this story from every outlet covering it into one summary — ranked by real-world impact, not just the latest headline — so you never miss what matters. What is BrevFeed? →

Key points

  • Compressed 1GB Wikipedia data to 105.39MB losslessly.
  • Uses a 163-million-parameter Transformer model as the sole predictor.
  • No hand-built compression models are used.
  • Requires two Nvidia B200 GPUs for compression/decompression.

Lossless Compression Achieved

Tufa Labs' tufazip, developed by researcher Tommy He, has successfully compressed 1GB of Wikipedia information into 105.39MB. This compression is lossless, meaning no data is lost during the process. The technique utilizes a 163-million-parameter Transformer model as its predictor.

Self-Learning Model Architecture

The tufazip package includes a 4.33MB pre-trained starter model, quantized for space efficiency, which is necessary for unpacking the compressed file. The full model is a 12-layer Transformer. The starter model contains two trained stages that are copied and learned independently to fill all 12 layers when the full model is active in memory. This system operates as a self-contained language model that learns dynamically.

Training Process and Determinism

The training occurs in two phases: an offline pre-training of the 34-million-parameter starter model, followed by online learning where the full model predicts each token and trains on it. The unpacking process mirrors these steps, rebuilding the model alongside the text. Predictor updates must be fully deterministic to ensure lossless compression.

Benchmark Performance and Distinction

Tufazip's approach distinguishes itself by not relying on traditional, hand-built compression models, unlike other top entries on the Large Text Compression Benchmark. It achieved the smallest result on this benchmark for any entry using only a neural network as its predictor. The current 105.39MB version ranked 13th on the benchmark as of November 19, while an earlier online-only version ranked second in September. This benchmark is specifically for text compression.

Research Focus and Hardware Requirements

This development is for research purposes and is not a commercial product. Running the compression or decompression process would require significant hardware, specifically two Nvidia B200 GPUs and two Intel Xeon Platinum 8570 CPUs, taking approximately one week to complete.

✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →

The daily brief

One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.

One email a day. Unsubscribe in one click, any time.

Today's brief

Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.

~4 min · 3 stories · Nov 21

▶ Play today's brief Listen on Spotify

New every morning, and the back catalogue is archived by date.

Reporting from

Tufa Labs' tufazip achieved lossless compression of 1GB of Wikipedia data to 105.39MB using a 163-million-parameter Transformer model as its sole predictor. This method is notable for not using any hand-built compression models and for its self-learning, on-the-fly adaptation, setting a new benchmark for neural network-only text compression.